In the background of the new era, the rapid development of networking and digitization has played a certain role in promoting the dissemination and development of opera culture. First, we normalize the opera action data of the recorded opera performers. Then, based on Laban’s movement theories, we propose a parallel one-dimensional convolutional network and recurrent neural network model that adds a way for attention to splice, encode, and decode the time sequences in the opera action data. Next, we employ the full connection layer to generate a sequence of Laban opera action symbols, enabling the recognition of Laban opera actions. Ultimately, we integrate the model with virtual reality technology to establish a virtual opera learning system. The model tests show that the suggested Laban opera action recognition and automatic generation method works better than the rule-based and template-matching methods for the same period. There is a significant difference (P<0.05) between the experimental group that used the opera virtual learning system and the control group, which was made up of opera fans who learned the old-fashioned way. This was seen in the post-test of technical quality, which looked at posture and foot movements. This paper’s designed system can enhance the interest and impact of opera enthusiasts in learning opera culture while also offering a reference path for the transmission of opera culture and related fields.